CN111241814B - Error correction method, device, electronic equipment and storage medium for speech recognition text - Google Patents

Abstract

The embodiment of the invention relates to the field of natural language processing, and discloses an error correction method and device for a voice recognition text, electronic equipment and a storage medium. In the invention, voice information is received; identifying at least one user intention corresponding to the voice information; according to the identified at least one user intention, selecting all data of the user intention from the cloud data set as a personalized fuzzy data set; and combining the personalized fuzzy data set and a preset basic fuzzy data set to correct the text identified according to the voice information, and reducing the data quantity required by error correction while ensuring the accuracy of error correction through the personalized user intention of the user, thereby improving the error correction efficiency.

Description

Speech recognition text error correction method, device, electronic device and storage medium

Technical Field

The embodiments of the present invention relate to the field of natural language processing, and in particular to a method, device, electronic device and storage medium for correcting errors in speech recognition text.

Background Art

With the development of artificial intelligence technology, users can control the operation of smart devices through voice. When controlling smart devices through voice, in order for smart devices such as speakers to more accurately recognize the user's voice and thus more accurately perform corresponding actions, it is necessary to convert the recognized user's voice information into text and correct the text. Current text error correction usually uses data in cloud data sets, detects the location of typos through language models, and corrects typos through pinyin sound-like features, stroke five-stroke edit distance features, and language model confusion features. A large amount of data usually needs to be added to the cloud data set to ensure that the language model can accurately identify typos in the text.

The inventors have discovered that there are at least the following problems in the related art: the amount of data in the cloud data set is too large, which will lead to too much error correction candidate data when correcting errors through the language model, thereby reducing the efficiency of error correction.

Summary of the invention

The purpose of the embodiments of the present invention is to provide a method, device, electronic device and storage medium for correcting speech recognition text, which can reduce the amount of data required for error correction while ensuring the accuracy of error correction through the user's personalized user intention, thereby improving the error correction efficiency.

To solve the above technical problems, an embodiment of the present invention provides a method for correcting speech recognition text, comprising: receiving speech information; identifying at least one user intent corresponding to the speech information; selecting all data of the user intent in a cloud data set as a personalized fuzzy data set based on the identified at least one user intent; and correcting the text recognized according to the speech information by combining the personalized fuzzy data set with a pre-set basic fuzzy data set.

An embodiment of the present invention also provides an error correction device for speech recognition text, comprising: a receiving module, a recognition module, a selection module, and an error correction module; the receiving module is used to receive speech information; the recognition module is used to recognize at least one user intention corresponding to the speech information; the selection module is used to select all data of the user intention in the cloud data set as a personalized fuzzy data set based on the recognized at least one user intention; the error correction module is used to combine the personalized fuzzy data set with a pre-set basic fuzzy data set to correct the text recognized according to the speech information.

An embodiment of the present invention also provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a method for correcting errors in speech recognition text.

The embodiment of the present invention further provides a storage medium storing a computer program, and the computer program, when executed by a processor, performs the above-mentioned method for correcting errors in speech recognition text.

Compared with the prior art, the embodiment of the present invention receives voice information, determines the user intention expressed in the voice information by identifying the voice information, for example, the received voice information expresses the intention of listening to music or reporting the weather, and uses the user intention to filter the data in the cloud data set, and filters out the data that may be used in this error correction process, that is, extracts the data in the cloud data set that has the same intention as the identified user intention as a personalized fuzzy data set, and combines the personalized fuzzy data set with the pre-set basic fuzzy data set to correct the text identified according to the voice information, thereby using the personalized user intention of the user contained in the voice information to reduce the amount of data used in the error correction process and improve the efficiency of error correction. In addition, since the set basic fuzzy data set contains basic confusing words such as homophones and homophonic words, the accuracy of error correction can be guaranteed by combining the preset basic fuzzy data set.

In addition, identifying at least one user intention corresponding to the voice information includes: identifying the voiceprint feature of the voice information; determining the user information corresponding to the voice information according to the voiceprint feature; obtaining historical voice information corresponding to the user information; determining at least one user intention that the user is used to according to the historical voice information, and using the at least one user intention that the user is used to as the at least one user intention corresponding to the voice information. In this way, the user's used user intention can be inferred from the user's historical voice information, thereby more accurately determining the user's personalized behavior.

In addition, determining at least one user intention that the user is accustomed to based on historical voice information includes: inputting features corresponding to the historical voice information into a pre-trained neural network model; wherein the neural network model is trained using features of the voice information of each user intention to identify the user's usage rate of each user intention; the feature value includes at least any one of the following features or a combination thereof: the total time of voice interaction corresponding to each user intention in the historical voice information, the user intention corresponding to the most recent voice information in the historical voice information, the user's age or the user's gender; determining at least one user intention that the user is accustomed to based on the output result of the neural network model.

In addition, before taking at least one user intention that the user is accustomed to as the at least one user intention corresponding to the voice information, it also includes: converting the voice information into text information, and identifying the intent of the text information to obtain the text intent; taking at least one user intention that the user is accustomed to as the at least one user intention corresponding to the voice information, including: if the text intent is the same as any user intention of the at least one user intention that the user is accustomed to, taking the at least one user intention that the user is accustomed to as the at least one user intention corresponding to the voice information; if the text intent is different from any user intention of the at least one user intention that the user is accustomed to, taking the at least one user intention that the user is accustomed to and the text intent together as the user intent corresponding to the voice information.

In addition, the text information is identified for intent to obtain the text intent, including: converting the text information into a vector matrix through a word embedding method; inputting the vector matrix into a pre-trained text classification model; and obtaining the text intent according to the output of the text classification model. The text intent is determined by the text information recognized by the currently received voice information, so that the determined personalized fuzzy data set contains personalized data that meets the current voice information, thereby making the error correction of the voice recognition text more accurate.

In addition, the personalized fuzzy data set and the pre-set basic fuzzy data set are combined to correct the text recognized based on the voice information, including: locating the position of the erroneous words in the text recognized by the voice information according to the personalized fuzzy data set and the pre-set basic fuzzy data set; selecting at least one replacement word for the erroneous word in the personalized fuzzy data set and the pre-set basic fuzzy data set; calculating the confusion score of at least one replacement word through a language model; and using the replacement word whose confusion score is less than a first preset threshold to correct the text recognized by the voice information.

In addition, the position of the erroneous word in the text recognized by the voice information is located, including: dividing the text recognized by the voice information into different segments; calculating the probability of each segment being an erroneous segment based on the correlation between the segments in the text; treating the erroneous segment with a probability greater than a second preset threshold as an erroneous word, and treating the position of the erroneous segment in the text as the position where the erroneous word appears in the text.

BRIEF DESCRIPTION OF THE DRAWINGS

One or more embodiments are exemplarily described by the pictures in the corresponding drawings, and these exemplary descriptions do not constitute limitations on the embodiments.

1 is a flow chart of a method for correcting errors in speech recognition text according to a first embodiment of the present invention;

2 is a flow chart of a method for correcting errors in speech recognition text according to a second embodiment of the present invention;

3 is a schematic diagram of the structure of a device for correcting speech recognition text according to a third embodiment of the present invention;

FIG. 4 is a schematic structural diagram of an electronic device according to a fourth embodiment of the present invention.

DETAILED DESCRIPTION

In order to make the purpose, technical scheme and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, it will be appreciated by those skilled in the art that in the embodiments of the present invention, many technical details are provided to enable the reader to better understand the present application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical scheme claimed in the present application can be implemented.

The following embodiments are divided for the convenience of description and shall not constitute any limitation on the specific implementation of the present invention. The embodiments may be combined with each other and referenced to each other without contradiction.

The first embodiment of the present invention relates to a method for correcting speech recognition text, comprising: receiving speech information; identifying at least one user intent corresponding to the speech information; selecting all data of the user intent in the cloud data set as a personalized fuzzy data set according to the identified at least one user intent; combining the personalized fuzzy data set with a pre-set basic fuzzy data set to correct the text recognized according to the speech information, and reducing the amount of data required for error correction while ensuring the accuracy of error correction through the user's personalized user intent, thereby improving the error correction efficiency. The implementation details of the speech recognition text correction method of this embodiment are described in detail below. The following content is only the implementation details provided for easy understanding and is not necessary for implementing this solution.

The specific process is shown in FIG1 . The first embodiment relates to a method for correcting errors in speech recognition text, including:

Step 101, receiving voice information. Specifically, when a user controls a smart device by voice, it is necessary to interact with a device such as a speaker by voice. The smart device receives the user's voice information through the speaker, for example, receiving voice information such as "play music" or "play novel" input by the user.

Step 102: Identify at least one user intention corresponding to the voice information.

Specifically, each voice message has its corresponding user intention. For example, the voice message "play music" corresponds to the user intention of playing music, the voice message "play novel" corresponds to the user intention of playing novel, and so on.

The method for identifying the user intent corresponding to voice information can be as follows: determine the user's usual user intent through the user's historical voice information, and use the user's usual user intent as the user intent corresponding to the identified voice information; determine the text intent through the text information recognized by the voice information, and use the text intent as the user intent corresponding to the identified voice information; and combine the user's usual user intent and text intent as the user intent corresponding to the voice information.

Step 103 , based on the identified at least one user intention, select all data of the user intention in the cloud data set as a personalized fuzzy data set.

Specifically, the cloud data set contains a large amount of data, and the data in the cloud data set can be classified according to user intentions. For example, the data of the user intention of playing music is put into one data list, and the data of the user intention of playing novels is put into another data list, so as to classify the data in the cloud data set according to user intentions. After identifying at least one user intention, the data list corresponding to the user intention is selected as a personalized fuzzy data set.

Step 104 , combining the personalized fuzzy data set with the preset basic fuzzy data set to correct the text recognized according to the voice information.

Specifically, the personalized fuzzy data set and the basic fuzzy data set are summarized. When using the personalized fuzzy data set and the basic fuzzy data set to correct the text information of speech information recognition, the Chinese word segmentation tools such as Jieba can be used for word segmentation first. For example, for the user interaction text "Come to Jay Chou's Shuang Jie Gun" recognized by the automatic speech recognition technology (ASR), the result after Jieba's word segmentation is "Come" | "A Song" | "Jay Chou" | "Of" | "Shuang" | "Jie Gun". After the text is segmented, the possible errors in the word segmentation are located, and the possible error positions can usually be located by using a language model based on word granularity. The language model is mainly processed by the N-Gram algorithm, and the position greater than the average score is taken as the possible error position.

The N-Gram mentioned above is an algorithm based on a statistical language model. Its basic idea is to perform a sliding window operation of size N on the content of the text according to bytes, forming a sequence of byte fragments of length N; each byte fragment is called a gram, and the frequency of occurrence of all grams is counted, and filtered according to a pre-set threshold to form a key gram list, which is the vector feature space of the text. Each gram in the list is a feature vector dimension; the model is based on the assumption that the appearance of the Nth word is only related to the previous N-1 words, but not to any other words, and the probability of the whole sentence is the product of the probability of occurrence of each word. These probabilities can be obtained by directly counting the number of times N words appear at the same time in the corpus. Commonly used algorithms are the binary Bi-Gram algorithm and the ternary Tri-Gram algorithm.

After locating the possible error positions, we select replacement words for the wrong words as error candidates in the personalized fuzzy dataset and the pre-set basic fuzzy dataset. For example, the possible error candidates for "爽" in the above description are "双", "霜", etc. For all possible situations, we need to process and evaluate all situations, and then sort the resulting confusion scores and select the most likely candidate. The confusion is still calculated based on the confusion score of the n-gram language model. Finally, the sentence with the lowest confusion score is compared with the set confusion score threshold. If the calculated result is less than the threshold, the corrected sentence is selected for output; otherwise, the original input text is selected for output.

Compared with the prior art, the embodiment of the present invention receives voice information, determines the user intention expressed in the voice information by identifying the voice information, for example, the received voice information expresses the intention of listening to music or reporting the weather, and uses the user intention to filter the data in the cloud data set, and filters out the data that may be used in this error correction process, that is, extracts the data in the cloud data set that has the same intention as the identified user intention as a personalized fuzzy data set, and combines the personalized fuzzy data set with the pre-set basic fuzzy data set to correct the text identified according to the voice information, thereby using the personalized user intention of the user contained in the voice information to reduce the amount of data used in the error correction process and improve the efficiency of error correction. In addition, since the set basic fuzzy data set contains basic confusing words such as homophones and homophonic words, the accuracy of error correction can be guaranteed by combining the preset basic fuzzy data set.

The second embodiment of the present invention relates to a method for correcting errors in speech recognition text. In the second embodiment of the present invention, the user's usual user intent is obtained through the user's historical speech information, and the text intent is determined through the text recognized by the currently received speech information, and the user's usual user intent and text intent are used to select data from the cloud data set as a personalized fuzzy data set. The specific process is shown in Figure 2, including:

Step 201, receiving voice information.

Step 202: Identify the voiceprint features of the speech information.

Step 203: determine the user information corresponding to the voice information according to the voiceprint feature.

Step 204: Acquire historical voice information corresponding to the user information.

Step 205: determining at least one user intention that the user is accustomed to based on the historical voice information.

Specifically, the voice information input by the user is recognized to identify the voiceprint features of the voice information. Since each user has his or her own unique voiceprint features, the user who input the voice information can be determined based on the voiceprint features of the voice information, thereby obtaining the historical voice information of the user.

When identifying user intent through text information, text classification (textCNN) technology can be used to classify text information. Text classification is implemented as follows: First, each word is converted into a vector of the same length through the word embedding method, and a vector matrix can be formed for a sentence. Secondly, the vector matrix is convolved through convolution, and the pooling layer and the fully connected layer are connected after the convolution layer, and finally the classification is performed through the logistic regression (softmax) layer.

When determining at least one user intent that the user is accustomed to using historical voice information, the above-mentioned text classification (textCNN) technology is used to establish a neural network model, and the neural network model is trained using the features of the voice information of each user intent to identify the user's usage rate of each user intent. The feature value includes at least any of the following features or a combination thereof: the total time of voice interaction corresponding to each user intent in the historical voice information, the user intent corresponding to the most recent voice information in the historical voice information, the user's age or the user's gender, etc. The feature value of the historical voice information is input into the neural network model, and at least one user intent that the user is accustomed to using can be determined through the output result of the neural network model. The output result can be: each user intent and the usage rate corresponding to each user intent. The user intent with a greater usage rate has a greater probability of being used by the user.

训练集中包括各个语音信息对应的特征值。由以上结果可知隐藏层第h个神经元的输入值为：The principle of the neural network model is explained below: The neural network model is mainly implemented through the BP neural network, and its main calculation formula is as follows: Since the BP neural network is a supervised learning algorithm, we specify the training data set of the model as D = {(x ₁ ,y ₁ ), (x ₂ ,y ₂ ), …, (x _q ,y _q )}, and the derivative of the activation function is known

The training set includes the feature values corresponding to each voice information. From the above results, we can know that the input value of the hth neuron in the hidden layer is:

其中，I_hh为隐藏层第h个神经元的输入值；W_ih为预先设置隐藏层的第h个神经元的权重；x_i为训练数据集中的特征值。

Among them, I _hh is the input value of the hth neuron in the hidden layer; _Wih is the pre-set weight of the hth neuron in the hidden layer; _Xi is the feature value in the training data set.

The output of the hth neuron in the hidden layer is: _Ohh =S( _{Ihh -Hh} ); _Ohh is the output value of the hth neuron in the hidden layer; _Hh is the preset threshold corresponding to the hth neuron in the hidden layer; S is the activation function in the above description.

I_oj为输出层第j个神经元的输入值；W_hj为预先设置的输出层第j个神经元的权重。The input value of the jth neuron in the output layer is:

I _oj is the input value of the j-th neuron in the output layer; W _hj is the preset weight of the j-th neuron in the output layer.

The output value of the j-th neuron in the output layer is: O _oj =S(I _oj -θ _j ); O _oj is the output value of the j-th neuron in the output layer; θ _j is the preset threshold corresponding to the j-th neuron in the output layer.

y_j为训练集中的特征值对应的用户意图。So the mean square error of the neural network model on (x _k ,y _k ) is:

_yj is the user intention corresponding to the feature value in the training set.

In order to minimize the total error E of the result, the model needs to update all parameters through continuous iteration. The BP neural network algorithm implements this process through the gradient descent algorithm. The gradient descent algorithm updates the negative gradient direction of the target result as the direction of parameter adjustment. For a given learning rate α, we can get:

化简之后可得：

After simplification, we can get:

减少预先设置隐藏层的神经元的权重；通过

减少隐藏层神经元对应的预设阈值。△W _hj = α*O _oj *(1-O _oj )*(y _j -O _oj )*O _hh , thereby reducing the error of the weight of each neuron. Similarly, the errors of other parameters in the neural network model can be corrected, such as reducing the preset threshold corresponding to the output layer neuron through △θ _j = -α*g _j ;

Reduce the weights of neurons in the pre-set hidden layer; by

Reduce the preset threshold corresponding to the hidden layer neurons.

Step 206, converting the voice information into text information, and identifying the intent of the text information to obtain the text intent. Specifically, after obtaining the user's usual user intent and text intent through the above neural network model, if the text intent is the same as any one of the at least one user intent that the user is accustomed to, the at least one user intent that the user is accustomed to is used as the at least one user intent corresponding to the voice information; if the text intent is different from any one of the at least one user intent that the user is accustomed to, the at least one user intent that the user is accustomed to and the text intent are used together as the user intent corresponding to the voice information.

Step 207, based on the user's usual user intent and text intent, select all the data of the user's intent in the cloud data set as a personalized fuzzy data set. Specifically, after obtaining the user's usual user intent and text intent, determine whether the intent set in the existing personalized fuzzy data set is the same as the user's usual user intent and text intent. If the intent set in the existing personalized fuzzy data set is the same as the user's usual user intent and text intent, there is no need to update the data in the existing personalized fuzzy data set; if the intent set in the existing personalized fuzzy data set is not the same as the user's usual user intent and text intent, delete the redundant intents in the existing personalized fuzzy data set, and add the missing intents in the existing personalized fuzzy data set, so as to update the existing personalized fuzzy data set. After the update, save the state of the current personalized fuzzy data set so that the personalized fuzzy data set can be compared again when the text of the speech information recognition is corrected next time.

Step 208, combining the personalized fuzzy data set with the preset basic fuzzy data set to correct the text recognized according to the voice information.

The step division of the above methods is only for the purpose of clear description. When implemented, they can be combined into one step or some steps can be split and decomposed into multiple steps. As long as they include the same logical relationship, they are all within the scope of protection of this patent; adding insignificant modifications to the algorithm or process or introducing insignificant designs without changing the core design of the algorithm and process are all within the scope of protection of this patent.

The third embodiment of the present invention relates to an error correction device for speech recognition text, as shown in Figure 3, comprising: a receiving module 31, a recognition module 32, a selection module 33, and an error correction module 34; the receiving module 31 is used to receive speech information; the recognition module 32 is used to identify at least one user intention corresponding to the speech information; the selection module 33 is used to select all data of the user intention in the cloud data set as a personalized fuzzy data set based on the identified at least one user intention; the error correction module 34 is used to combine the personalized fuzzy data set with a pre-set basic fuzzy data set to correct the text recognized according to the speech information.

It is not difficult to find that this embodiment is a system embodiment corresponding to the first embodiment, and this embodiment can be implemented in conjunction with the first embodiment. The relevant technical details mentioned in the first embodiment are still valid in this embodiment, and in order to reduce repetition, they are not repeated here. Accordingly, the relevant technical details mentioned in this embodiment can also be applied in the first embodiment.

In addition, the recognition module 32 is used to recognize the voiceprint features of the voice information; determine the user information corresponding to the voice information based on the voiceprint features; obtain historical voice information corresponding to the user information; determine at least one user intention that the user is accustomed to based on the historical voice information, and use the at least one user intention that the user is accustomed to as the at least one user intention corresponding to the voice information.

In addition, the recognition module 32 is used to input the features corresponding to the historical voice information into a pre-trained neural network model; wherein the neural network model is trained using the features of the voice information of each user intention, and is used to identify the user's usage rate of each user intention; the feature value includes at least any one of the following features or a combination thereof: the total time of voice interaction corresponding to each user intention in the historical voice information, the user intention corresponding to the most recent voice information in the historical voice information, the user's age or the user's gender; and at least one user intention that the user is accustomed to is determined based on the output result of the neural network model.

In addition, the recognition module 32 is used to convert the voice information into text information, and recognize the intent of the text information to obtain the text intent.

In addition, the recognition module 32 is used to convert text information into a vector matrix through a word embedding method; input the vector matrix into a pre-trained text classification model; and obtain the text intent based on the output result of the text classification model.

In addition, the error correction module 34 is used to locate the position of erroneous words in the text recognized by the speech information based on the personalized fuzzy data set and the preset basic fuzzy data set; select at least one replacement word for the erroneous word in the personalized fuzzy data set and the preset basic fuzzy data set; calculate the confusion score of at least one replacement word through the language model; and use the replacement word whose confusion score is less than the first preset threshold to correct the text recognized by the speech information.

It is worth mentioning that all modules involved in this embodiment are logic modules. In practical applications, a logic unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. In addition, in order to highlight the innovative part of the present invention, this embodiment does not introduce units that are not closely related to solving the technical problem proposed by the present invention, but this does not mean that there are no other units in this embodiment.

A fourth embodiment of the present invention relates to an electronic device, as shown in FIG4 , comprising at least one processor 401; and a memory 402 communicatively connected to the at least one processor 401; wherein the memory 402 stores instructions executable by the at least one processor 401, and the instructions are executed by the at least one processor 401 so that the at least one processor 401 can execute the above-mentioned speech recognition text error correction method.

The memory 402 and the processor 401 are connected in a bus manner, and the bus may include any number of interconnected buses and bridges, and the bus connects various circuits of one or more processors 401 and the memory 402 together. The bus can also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and are therefore not further described herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be one element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices on a transmission medium. The data processed by the processor is transmitted on a wireless medium via an antenna, and further, the antenna also receives data and transmits the data to the processor 401.

The processor 401 is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management and other control functions. The memory 402 can be used to store data used by the processor 401 when performing operations.

A fifth embodiment of the present invention relates to a computer-readable storage medium storing a computer program, which implements the above method embodiment when executed by a processor.

That is, those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, and the program is stored in a storage medium, including a number of instructions to enable a device (which can be a single-chip microcomputer, chip, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

Those skilled in the art will appreciate that the above embodiments are specific embodiments for implementing the present invention, and that in actual applications, various changes may be made in form and detail without departing from the spirit and scope of the present invention.

Claims (9)

1. A method for error correction of speech recognition text, comprising:

receiving voice information;

identifying at least one user intention corresponding to the voice information;

according to the identified at least one user intention, selecting all data of the user intention from a cloud data set as a personalized fuzzy data set;

combining the personalized fuzzy data set with a preset basic fuzzy data set to correct the text identified according to the voice information;

the identifying at least one user intention corresponding to the voice information comprises the following steps:

identifying voiceprint features of the voice information;

determining user information corresponding to the voice information according to the voiceprint characteristics;

acquiring historical voice information corresponding to the user information;

and determining at least one user intention of the user according to the historical voice information, and taking the at least one user intention of the user as the at least one user intention corresponding to the voice information.

2. The method of claim 1, wherein said determining at least one user intention of a user from said historical speech information comprises:

inputting the characteristic value corresponding to the historical voice information into a pre-trained neural network model; the neural network model is trained by utilizing characteristic values of voice information of each user intention and is used for identifying the utilization rate of the user on each user intention;

the characteristic value at least comprises any one or combination of the following characteristics: the total time of voice interaction corresponding to each user intention in the historical voice information, the user intention corresponding to the last voice information in the historical voice information, the age of the user or the gender of the user;

and determining at least one user intention used by the user according to the output result of the neural network model.

3. The error correction method of a speech recognition text according to claim 1 or 2, characterized by further comprising, before said regarding at least one user intention familiar to said user as at least one user intention corresponding to said speech information:

converting the voice information into text information, and carrying out intention recognition on the text information to obtain text intention;

the step of using the at least one user intention which is used by the user as the at least one user intention corresponding to the voice information comprises the following steps:

if the text intention is the same as any user intention of at least one user intention which is used by the user, taking the at least one user intention which is used by the user as at least one user intention corresponding to the voice information;

and if the text intention is different from any user intention of at least one user intention which is used by the user, using the at least one user intention which is used by the user and the text intention together as the user intention corresponding to the voice information.

4. The method for text error correction in speech recognition according to claim 3, wherein said identifying the text information for intent results in text intent, comprising:

converting the text information into a vector matrix by a word embedding method;

inputting the vector matrix into a pre-trained text classification model;

and obtaining the text intention according to the output result of the text classification model.

5. The method of claim 1, wherein said combining the personalized fuzzy data set with a pre-set basic fuzzy data set to correct the text recognized from the speech information comprises:

positioning the position of the error word in the text identified by the voice information according to the personalized fuzzy data set and a preset basic fuzzy data set;

selecting at least one replacement word of the wrong word from the personalized fuzzy data set and a preset basic fuzzy data set;

calculating confusion degree scores of the at least one replacement word through a language model respectively;

and correcting the error of the text identified by the voice information by using the replacement words with the confusion degree score smaller than a first preset threshold value.

6. The method for correcting errors in speech recognition text according to claim 5, wherein locating the position of the erroneous word in the text recognized by the speech information comprises:

dividing the text identified by the voice information into different word segments;

according to the correlation among the word segments in the text, respectively calculating the probability that each word segment is an error word segment;

and taking the error word segment with the probability larger than a second preset threshold value as the error word, and taking the position of the error word segment in the text as the position of the error word in the text.

7. An error correction device for speech recognition text, comprising: the device comprises a receiving module, an identification module, a selection module and an error correction module;

the receiving module is used for receiving voice information;

the recognition module is used for recognizing voiceprint features of the voice information; determining user information corresponding to the voice information according to the voiceprint characteristics; acquiring historical voice information corresponding to the user information; determining at least one user intention of the user according to the historical voice information, and taking the at least one user intention of the user as at least one user intention corresponding to the voice information;

the selecting module is used for selecting all data of the user intention from a cloud data set as a personalized fuzzy data set according to the identified at least one user intention;

the error correction module is used for correcting errors of texts identified according to the voice information by combining the personalized fuzzy data set and a preset basic fuzzy data set.

8. An electronic device, comprising:

at least one processor; the method comprises the steps of,

a memory communicatively coupled to the at least one processor; wherein,,

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the error correction method of speech recognition text as claimed in any one of claims 1 to 6.

9. A computer readable storage medium storing a computer program, wherein the computer program when executed by a processor implements the error correction method of speech recognition text according to any one of claims 1 to 6.

Images